Token efficiency for Facebook automation/scrapping
Hey everyone — been running OpenClaw on a Mac mini M4 for a few weeks and loving the concept, but I'm getting destroyed on API costs and need some advice.
The setup:
- Scheduled cron task fires at 7am daily
- Agent uses Chrome browser relay to scan 4 Facebook groups for construction leads
- Takes 1-3 snapshots per group to read posts
- Reads/writes a JSON dedup file
- Sends results via Telegram
- Model: Claude Sonnet (trying to move to Haiku)
The problem:
One scan run is costing me $5-6 every time. I just watched it burn $5.57 in a single run. For what should be a simple read-and-report task, that's insane.
What I suspect:
Full page snapshots of Facebook are dumping massive amounts of DOM/HTML into the context. 4 groups × 2-3 snapshots = potentially hundreds of thousands of tokens per run. I also don't know if conversation history is being carried across tool calls within the same session.
My questions:
1. Is there a way to limit what a snapshot captures — text only, visible content only, etc.?
2. Is there a context trimming or transcript hygiene setting I should be using for scheduled tasks?
3. Does each cron-triggered task start with a fresh context or does it carry history?
4. Has anyone solved the browser snapshot token problem for scraping-style tasks?
Happy to share my full skill/prompt if helpful. Just trying to get this under $0.50 a run where it should be. Thanks
The setup:
- Scheduled cron task fires at 7am daily
- Agent uses Chrome browser relay to scan 4 Facebook groups for construction leads
- Takes 1-3 snapshots per group to read posts
- Reads/writes a JSON dedup file
- Sends results via Telegram
- Model: Claude Sonnet (trying to move to Haiku)
The problem:
One scan run is costing me $5-6 every time. I just watched it burn $5.57 in a single run. For what should be a simple read-and-report task, that's insane.
What I suspect:
Full page snapshots of Facebook are dumping massive amounts of DOM/HTML into the context. 4 groups × 2-3 snapshots = potentially hundreds of thousands of tokens per run. I also don't know if conversation history is being carried across tool calls within the same session.
My questions:
1. Is there a way to limit what a snapshot captures — text only, visible content only, etc.?
2. Is there a context trimming or transcript hygiene setting I should be using for scheduled tasks?
3. Does each cron-triggered task start with a fresh context or does it carry history?
4. Has anyone solved the browser snapshot token problem for scraping-style tasks?
Happy to share my full skill/prompt if helpful. Just trying to get this under $0.50 a run where it should be. Thanks
The AI that actually does things. Emails, calendar, home automation - all from your favorite chat app. New shell, ready to help. The lobster way.
134,827Members
Resources
Similar Threads
Was this page helpful?
